Download Speculation in Commodity Futures Markets

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

1973 oil crisis wikipedia , lookup

Transcript
Speculation in Commodity Futures Markets, Inventories and the
Price of Crude Oil
Sung Je Byun
∗
†
March 1, 2016
Abstract
Refiners have an incentive to hold inventories even if they anticipate falling crude oil
prices. This paper develops a model of the convenience yield arising from holding inventories. Using data on inventories and futures contract prices, I show that the convenience
yield is inversely related to inventories, and positively related to oil prices. In addition to
exhibiting seasonal and procyclical behaviors, I show that the historical convenience yield
averages about 19% of the oil price from March 1989 to November 2014. Although some
have argued that a breakdown of the relationship between crude oil inventories and prices
following increased participation by financial investors after 2003 was evidence of an effect
of speculation, I find that the proposed price-inventory relationship is stable over time. In
light of this evidence, I conclude that the contribution of financial investors’ activities to the
crude oil market is weak.
*JEL classification: G13, G15, G17, G23
*Keywords: Speculation, Convenience yield, Forecasting oil prices, Stable oil price-inventory
relationship
∗
I am very grateful to James Hamilton, Allan Timmermann, Valerie Ramey, Alexis Toda, Thomas Baranga,
Christiane Baumeister, Soojin Jo and anonymous referees for their helpful comments and suggestions on earlier
drafts of this paper.
†
Financial Industry Studies Department, Federal Reserve Bank of Dallas, 2200 N. Pearl Street, Dallas, TX,
USA. The views expressed are those of the author and do not necessarily reflect the views of the Federal Reserve
Bank of Dallas or the Federal Reserve System.
1
1
Introduction
The recent volatility of crude oil prices has renewed interest in the behavior of crude oil inventories. This paper examines the role of inventories in refiners’ gasoline production and develops
a structural model of the relation between crude oil prices and inventories.
In a competitive commodity market, a producer makes the optimal storage decision by
equating his expected benefit with the relevant cost of holding inventories. The positive benefit
could motivate a producer to hold inventories even if the producer anticipates falling prices in
the future. Given this motivation, the importance of inventories for storable commodities has
been widely recognized in the theory of storage literature. Among early researchers, Brennan
(1958) defines this benefit as the convenience yield, which is the flow of services that accrues to
an owner of the physical commodity but not to an owner of a contract for future delivery of the
commodity.
In this paper, I propose an equilibrium storage model of the global crude oil market to study
the role of crude oil inventories in refiners’ gasoline production and to explain fluctuations of
crude oil prices in terms of risk-averse refiners’ benefits to hold inventories. First, I model determinants of crude oil inventories, building upon earlier research by Eckstein and Eichenbaum
(1985) and Considine (1997). Specifically, I model the role of crude oil inventories directly as enhancing refiners’ gasoline production by treating inventories as an essential factor of production
following Kydland and Prescott (1982, 1988) and Ramey (1989). I derive an inverse relationship
between convenience yield and the level of crude oil inventories, which has never been done in
the literature. Instead, the common approach as in Pindyck (1994, 2001), Alquist and Kilian
(2010) and Knittel and Pindyck (2015) is to assume such an inverse relationship when modeling
determinants of crude oil prices and to infer convenience yield from the term spread observed
in prices of crude oil futures contracts at two different maturities. Second, I provide empirical
evidence that the convenience yield is inversely related to crude oil inventories, yet positively
related to crude oil prices. Moreover, it exhibits a procyclical behavior along the business cycle
with the convenience yield being increased during the expansionary period. In general, these
findings are consistent with the earlier conjecture in the theory of storage. Third, I find from
model estimates that the convenience yield accounts to about 19% of oil price on average from
March 1989 to November 2014, holding other factors of production constant. Regarding the
2
observed seasonality in oil inventories, I find a higher convenience yield for spring and summer than for other seasons. These complement earlier researchers’ empirical findings. Fourth,
I identify a structural change in crude oil market fundamentals since 2004 from a rising permanent component of crude oil prices. Such a structural break is also evidenced in changing
model parameter estimates for before and after June 2004, which is consistent with Tang and
Xiong (2012), Büyükşahin and Robe (2014) and Hamilton and Wu (2014). Lastly, I show that
the proposed model provides more accurate 3-, 6-, 9- and 12-month-ahead crude oil spot price
forecasts than the random walk model, indicating the stable price-inventory relation despite the
potential structural break.
I illustrate one use of the proposed model by re-examining the possible contribution of
financial investors on the stable price-inventory relation in the crude oil market. Given rapidly
increasing financial investors’ activities since 2004 (Irwin and Sanders 2011), a public debate
has centered on the potential contribution of financial investors’ participation in the commodity
futures markets (Masters 2008, Kennedy 2012). One piece of evidence that may lead to such
perception was provided by Merino and Ortiz (2005), who noted that the traditional model
of the relation between crude oil prices and inventories, such as that developed by Ye et al.
(2002), broke down after financial investor participation in crude oil markets increased. Merino
and Ortiz extended this model to include a contribution of the long positions of non-commercial
traders, and they documented that such a specification could better explain the relation between
prices and inventories over 2001-2004.
To investigate the potential contribution of financial investors’ activities on the crude oil
market, I consider previous forecasting models provided by Ye et al. (2002) and Merino and
Ortiz (2005) for two reasons. First, these models are grounded in the fundamental relationship
between the two variables. Second, such simple models provide a good framework for examining
popular political perceptions on the crude oil market. In this paper, I reproduce the good
in-sample fit of the Ye specification, its breakdown after 2001 and the improvement provided
by non-commercial positions over 2001-2004. However, I also find that even the latter relation
falls apart after 2004. I show that although the traditional model seems to do well over the
period 1992-2001, it is in fact misspecified even over that period and simply appears to do
well because there is relatively little change in the randomly driven component of crude oil
prices over that period. Despite the structural change after 2004, I show that the model of the
3
inventory-price relation proposed here is stable over time and furthermore, can account for both
the apparent success of the Ye et al. (2002) specification on earlier data as well as its breakdown
on subsequent data. This finding is consistent with Kilian and Murphy (2014) who document
the stable forecasting relationship between the oil market variables using a structural vector
autoregressive (VAR) model. Although the approach taken here is indirect by focusing on the
refineries’ storage decision, this paper adds further support to the recent academic literature
documenting empirical evidence for no contribution of financial investors (or limited at most)
from various angles (Fattouh et al. 2013, Kilian and Murphy 2014, Kilian and Lee 2014, Juvenal
and Petrella 2015, Knittel and Pindyck 2015, Hamilton and Wu 2015).
The remainder of the paper proceeds as follows. Section 2 introduces a storage model where
the convenience yield arises from the refiners’ expected benefits to hold crude oil inventories.
After introducing ways to deal with empirical issues, I provide a convenience yield measure at
the end of section 2. In section 3, I begin by introducing data and an empirical procedure
for estimating the proposed storage model. Next, I document estimation results together with
explanations for maximum likelihood estimates of structural model parameters as well as empirical properties of the proposed convenience yield. After providing evidence on the structural
break since 2004, I show the stable price-inventory relation proposed in this paper is robust to
the structural break in the crude oil market. In section 4, I illustrate one use of the proposed
model by investigating earlier forecasting models and show the strong forecasting relationship
of inventories and prices, indicating a weak contribution of financial investors’ activities in the
crude oil market. Lastly, I conclude with some remarks.
2
Theoretical Model
In this section, I propose an equilibrium storage model of the global crude oil market to study
the role of crude oil inventories in refiners’ gasoline production. I study the optimal storage
decision of refiners who have an incentive to hold inventories despite anticipated falling crude
oil prices. Note that the first-order conditions for refiners have to hold regardless of the actions
of financial investors in futures markets, because a refiner can always respond to arbitrage
opportunities from storing oil and selling a futures contract. This is separate from the behavior
of financial investors, who roll over maturing futures contracts forward rather than taking a
4
physical delivery of crude oil (Smith 2009, Hamilton and Wu 2015). However, financial investors’
activities could be reflected through risk premiums that the refinery pays for its hedging activities
in the crude oil market (Hamilton and Wu 2014). Although financialization in commodity
markets affects risk-premiums, my model differs from Basak and Pavlova (2015). There is no
role except consumption smoothing for storable commodities in their model, which neglects the
role of storable commodities in the theory of storage literature. Furthermore, the objective in
this paper is to build a structural relationship between relevant oil market variables and to
explain fluctuations of crude oil prices in terms of risk-averse refiners’ benefits to hold inventory,
i.e., convenience yield. From this perspective, the adopted approach differs from Knittel and
Pindyck (2015), in which the authors seek to isolate the speculative component of changes
in prices and in the convenience yield. Lastly, the proposed model provides a fully specified
structural relationship among oil market variables following the evidence established in the
literature. This differs from the approach in the recent structural VAR literature such as Kilian
and Murphy (2014), Kilian and Lee (2014) and Juvenal and Petrella (2015), whose objectives
are to disentangle various sources of price changes using residuals obtained from an unrestricted
VAR model for oil market variables.
2.1
Cost Function
Consider a representative producer who purchases a quantity qtP of crude oil at price St per
barrel, of which qtU is used up in current production of a consumption good such as gasoline and
the remainder goes to increase inventories it :1
it = it−1 + qtP − qtU ,
(1)
where qtP − qtU corresponds to net additional crude oil inventories during the time period t. If
the quantity of oil consumed by the representative producer (qtU ) is smaller than the quantity
purchased (qtP ), inventories accumulate and vice versa.
1
The proposed model abstracts a refiner’s decision for gasoline production using crude oil from its complicated
production decisions for multiple intermediary goods. It is because crude oil itself has accounted for more than
40% of the wholesale gasoline price. Excluding sales taxes, costs for the crude oil resource dominate the refinery’s
remaining operation costs such as manufacturing and marketing. Given its importance in gasoline production, I
focus on crude oil inventories and explain the benefit from holding raw material inventories. From this perspective,
the approach in this paper is different from most earlier literature that studies roles of finished good inventories
in smoothing a production schedule and reducing marketing costs.
5
The cost function summarizes both the production and storage technology for the representative producer. Given the current crude oil spot price (St ), the firm’s costs come from two
components: costs for purchasing resources and those for storing inventories over one period.
The former is the product of crude oil spot price and amounts purchased (qtP ), and the latter
is assumed to be proportional to the current crude oil spot price.2 For reflecting the idea of
limited storage facilities in the short-run, storage costs are further assumed to take the form of
a convex quadratic function. Given the current crude oil spot price (St ) and previous level of
inventories (it−1 ), the representative producer’s cost function becomes
c1 c qtU , it ; it−1 , St = St qtU + (it − it−1 ) + St c0 + i2t ,
2
where I plug in the accounting identity of inventories following (1) after rearranging it for a
quantity purchased (qtP ) and c0 , c1 > 0.
2.2
Production Function
The importance of inventories has been widely recognized in the theory of storage developed by
Kaldor (1939), Working (1949) and Brennan (1958). In particular, Brennan (1958) defines the
convenience yield as the flow of services that accrues to an owner of the physical commodity
but not to an owner of a contract for future delivery of the commodity. These services can arise
from reducing the probability of a stock-out of inventories, from inventories’ role in production
smoothing, and from future production cost saving. For example, Pindyck (1994) describes
the role of inventories in production is “to reduce costs of adjusting production and to reduce
marketing costs by facilitating schedule and avoiding stockouts”. With his proposed measure
for the intangible convenience yield using available futures contract prices, he finds evidence
for such roles for copper, heating oil, and lumber. Further empirical tests for the convenience
yield in commodities include Fama and French (1987, 1988), Deaton and Laroque (1992), Ng
and Pirrong (1994), Routledge et al. (2000), Bryant et al. (2006), Gorton et al. (2013), Acharya
et al. (2013), and Pirrong (2011).
The role of inventories as a factor of production has been widely adopted in earlier litera2
The cost for storage is defined in a broad sense: it includes the cost for insurance and transportation beside
that for using storage facilities. In general, these costs are determined as being proportional to the value that is
stored.
6
ture. For example, Kydland and Prescott (1982, 1988) and Ramey (1989) treat inventories as
essential factors of production for studies of aggregate fluctuations in the U.S. economy. Focusing on the petroleum refining industry, Considine (1997) examines determinants of inventory
investment under joint production by treating inventories as quasi-fixed factors of production.
In this paper, I model crude oil inventories as directly increasing the refinery’s production capabilities. To motivate this role of inventories further, consider the refinery’s decision problem
for a periodic production schedule under resource constraints. Compared to firms in other industries, the refinery’s production heavily depends on the raw material, in this case, crude oil.
It generally takes more than a few weeks to receive additional crude oil delivery after placing
an order in the spot market or making a transaction in the derivative market. For its current
gasoline production, the refinery can use, at most, crude oil resources that are either carried over
from the previous period or are delivered currently from the past transaction. In response to
uncertainties in the aggregate gasoline demand and petroleum prices, the available inventories in
the refinery’s storage determine the attainable level of production and can improve production
efficiency without needing to make costly production adjustments. Hence, the refinery has a
motivation for holding inventories despite anticipated falling crude oil prices in the future. Given
the realized technology process (zt ), I propose a production function of the form,
o
α n
f qtU ; it−1 , zt = ezt qtU
1 − e−θt ·it−1 ,
where α ∈ (0, 1) is the output share of the crude oil resource and θt > 0 is the utilization
parameter governing the production function’s dependence on the previous level of inventory.
The subscript t on utilization parameter (θt ) is used to allow the possiblity for taking different
values depending on the season of the year - for example, taking a smaller value when the refinery
tries to produce more gasoline for the summer driving season. The utilization function denoted
by the term in curly brackets is bounded between 0 and 1, which is a concave function of the
previous level of inventory. This captures the nonlinearity of a refinery’s production arising
from inventory stock-outs, approaching 0 at an increasing rate for it−1 being close to 0 and
approaching 1 at a reducing rate for sufficiently large it−1 . The resource-augmenting technology
process (zt ) follows the random walk process, that is, zt+1 = zt + 1,t+1 with 1,t+1 ∼ N 0, σ12 .
7
2.3
Theoretical Model
With the cost function and the production function as introduced earlier, the representative
producer faces a dynamic programming problem. At the beginning of each period t, the representative producer faces the crude oil spot price of St , the realized technology process zt and the
exogenously determined real interest rate rt . Given previous inventory (it−1 ), he makes decisions
on the resource demand (qtU ) and inventories (it ) jointly in order to maximize his lifetime profits.
Suppose he is a price taker in the crude oil market and he is risk averse with discount factor
(Λt ),3 which governs his risk aversion. Suppressing the superscript in the resource demand (qt ),
the representative producer’s objective is thus to choose {qt , it }∞
t=1 so as to maximize
"
Π = max E0
{qt ,it }
∞ Y
t
X
#
Λτ e
−rτ
{f (qt ; it−1 , zt ) − c (qt , it ; it−1 , St )} ,
t=1 τ =1
where i−1 is given. Note, I pose this as the representative agent’s problem of producing consumption goods rather than gasoline, for the purposes of studying the role of inventories in a
broad perspective - facing global phenomena of acculmulating crude oil inventories, or petroleum
as a whole. In spite of this subtle difference, I use the term “refinery” interchangeably for the
facile understanding of our theoretical model and its implication.
The optimality conditions for the representative producer are summarized as:
α exp (αzt ) qtα−1 {1 − exp (−θt it−1 )} = St ,
(2)
α
Et Λt+1 θt+1 exp (−rt+1 − θt+1 it + αzt+1 ) qt+1
= St (1 + c1 it ) − Et [Λt+1 exp (−rt+1 ) St+1 ] . (3)
Equation (2) is the optimality condition associated with the refinery buying one more barrel
of crude oil and using the crude oil immediately for its current production. The resource demand is determined where the marginal product equals the marginal cost. Equation (3) is the
optimality condition for the storage decision, in which the refinery equates expected marginal
benefits with marginal costs of holding an additional barrel of crude oil inventory. A positive
value of the expected marginal benefits, left-hand side in (3), introduces the convenience yield
as directly increasing the refinery’s future production capabilities. The representative producer
β·U 0 (cτ +1 )
In a standard macro model, we have Λτ = U 0 (cτ ) erτ , where U (·) represents the agent’s utility function,
β is the discount factor and rτ is the interest rate.
3
8
has an incentive of holding a positive level of inventory despite an expected capital loss in the
future.
2.4
Aggregation, Seasonality and the Crude Oil Spot Price
The theory presented so far applies to one representative producer; however, available observations are aggregated at the global level. Given the trend growth observed from the global crude
oil production, consumption and inventories, we assume that the number of representative producers increases over time at a fixed rate. Let Nt be the number of representative producers
at period t and g be the exogenously determined average growth rate of the representative producer, i.e. Nt = (1 + g) Nt−1 . The aggregate resource demands (Qt ) and inventories (It ) become
Qt = Nt · qt and It = Nt · it , where Qt and It will be associated with globally observed aggregate
quantities of the crude oil consumed and inventories. Assuming competitiveness in the global
refinery industry following Eckstein and Eichenbaum (1985) and Pindyck (2001), the aggregate
production F and cost functions C become F (·) = Nt · f (·) and C (·) = Nt · c (·) where the
representative producer’s production f (·) and the cost function c (·) are as defined earlier.
Next, I introduce seasonally varying utilization parameters to deal with the strong seasonality
observed both in the crude oil consumption and inventories. Specifically, I conjecture that
seasonal variations exist in the representative producer’s benefit, accordingly, the seasonally
varying convenience yield. In the northern hemisphere, for example, the aggregate demands for
gasoline increase during the summer for traveling. In most cases, refiners are able to predict
these seasonally varying aggregate demands and tend to accumulate inventories in advance when
it is profitable to meet increasing demands for gasoline by utilizing their production facilities
efficiently. I propose to capture seasonal variations of the utilization parameters as, θt equals θ1
if the month of the time period t corresponds to March, April, May, θ2 for June, July, August,
θ3 for September, October, November, and θ4 for December, January, February.
Lastly, I model the crude oil spot price process following the long-term/short-term model
of Schwartz and Smith (2000) because of the model’s flexibility to capture observed behaviors
in the crude oil prices. In this empirical application, I adapt their model to a discrete time
stochastic process to obtain the closed-form forecasts for the future crude oil spot price.
Suppose that the logarithm of crude oil spot price (log St ) consists of the long-term trend
9
(ξt ) and the short-term deviation (χt ).4 The long-term trend follows the random walk process,
reflecting the idea that oil prices themselves behave like a random walk at each time period. On
the other hand, the short-term deviation from the long-term trend follows the mean-reverting
AR(1) process:
ln St+1 ≡ ξt+1 + χt+1 ,
ξt+1 = ξt + 2,t+1 ,
χt+1 = k · χt + 3,t+1 ,
where k ∈ (−1, 1) is the mean-reversion parameter. 2,t+1 and 3,t+1 are normally distributed
innovation processes with E [2,t+1 ] = E [3,t+1 ] = 0, V ar [2,t+1 ] = σ22 , V ar [3,t+1 ] = σ32 for ∀t.
Though not serially uncorrelated with their own processes, innovations are correlated to each
other only at the same time period, i.e., corr [2,t+1 , 3,s+1 ] = ρ for ∀t = s.
Given the crude oil spot price process, I adopt the risk-neutral valuation framework to
solve the representative producer’s forecasting problem in the storage decision (3) as if he is
risk-neutral.5 Specifically, I specify risk-neutral processes by subtracting risk premiums from
underlying stochastic processes of state variables (zt , ξt , χt ). While risk premiums are equilibrium
prices for risks the producer pays for his hedging activities in the crude oil market, this form
of risk adjustment is frequently adopted in the literature (Schwartz and Smith 2000, Pirrong
h
i0
ez , λ
eξ, λ
eχ risk premiums, I assume zero,
2011, and Hamilton and Wu 2014). Denoting by Υ ≡ λ
constant, and linear state-dependent risk premiums on the technology (zt ), long-term trend (ξt )
ez ≡ 0, λ
eξ = λξ , λ
eχ ≡ λχ + $χt with
and short-term deviation processes (χt ) respectively: λ
4
Another way to understand the adopted spot price process is a two-factor model. While the long-term factor
evolves according to a random walk process, the short-term factor reflects a temporal deviation from the first
factor. In this paper, the level of crude oil inventories, accordingly the convenience yield, helps disentangle the
long- and short-term factors from the observed crude oil futures prices of which risk premiums are adjusted by
the risk-neutral valuation framework.
5
Implicit assumptions are the deterministic interest rate and the redundancy of the futures contract. The
former guarantees the price equivalance between futures and forward contracts. More importantly, the latter validates the proposed approach of specifying risk-neutral processes by subtracting the risk premiums from underlying
processes. See more details for the application of the risk-neutral valuation framework in Duffie (Dynamic Asset
Pricing Theory, 2001, pp. 167-174)
10
λξ > 0, λχ > 0 and $ being constants. Hence, risk-neutral stochastic processes are of the form:


 zt+1

 ξ
 t+1

χt+1
  0   1 0 0
 
 
 =  −λ  +  0 1 0
ξ 
 

 
 
−λχ
0 0 kQ






Q
1,t+1
  zt  

 
  ξ  +  Q
  t   2,t+1

 
χt
Q
3,t+1



,


Q
Q
where k Q ≡ k − $ ∈ (−1, 1). Assuming that Q
1,t+1 is independent of both 2,t+1 and 3,t+1 ,
Q
Q
innovations for risk-neutral stochastic processes (Q
1,t+1 , 2,t+1 , 3,t+1 ) have identical properties
as described earlier.
Since the crude oil spot price is conditionally log-normally distributed, the closed-form forecasting for the one-period-ahead crude oil spot price is provided as6
EtQ [St+1 ]
1 Q
= exp ξt + k χt − λξ − λχ + s1 ,
2
Q
(4)
Q
2
2
where sQ
1 ≡ V art [log St+1 ] = σ2 +σ3 +2ρσ2 σ3 . Here the superscript Q indicates the calculation
under the risk-neutral processes as opposed to those under the physical measure such as Et [·]
and V art [·].
Similarily, the τ -periods-ahead crude oil spot price forecast is inferred recursively by
EtQ [St+τ ]
1
Q
= exp
+ V art [log St+τ ]
(5)
2
τ
τ −1 1 Q
= exp ξt + k Q · χt − τ · λξ − λχ 1 + k Q + · · · + k Q
+ sτ ,
2
2
where sQ
τ ≡ τ · σ2 +
EtQ [log St+τ ]
1−(kQ )
2τ
1−(kQ )2
τ
· σ32 + 2ρσ2 σ3 ·
1−(kQ )
1−kQ
is the conditional variance associated with
τ -periods-ahead forecast of the crude oil spot price.
6
The conditional expectation and the variance of log St+1 under the risk-neutral processes are respectively given
as, EtQ [log St+1 ] = ξt + kQ χt − λξ − λχ and V artQ [log St+1 ] = σ22 + σ32 + 2ρσ2 σ3 . Hence, one-month-ahead crude
oil spot price forecast becomes EtQ [St+1 ] = exp EtQ [log St+1 ] + 21 V artQ [log St+1 ] due to Jensen inequality. For
h
i
n
o
arbitrary φ > 0, it can be further generalized as EtQ (St+1 )−φ = exp −φ · ξt + kQ · χt − λξ − δ + 12 φ2 sQ
1 .
11
2.5
Equilibrium Prediction
With approaches introduced in the previous subsection, the producer’s optimality conditions in
(2) and (3) become
α exp (αzt ) (Qt /Nt )−(1−α) {1 − exp (−θt It−1 /Nt−1 )} = exp (ξt + χt ) ,
αφ θt+1 (1 + g) {1 − exp (−θt+1 It /Nt )}φ exp (−θt+1 It /Nt )
1 2 Q
Q
2
× exp −rt+1 + φ zt − ξt − k χt + λξ + λχ + φ s1 + σ1
2
1 Q
Q
= exp (ξt + χt ) 1 + c1 It /Nt − exp −rt+1 − 1 − k χt − λξ − λχ + s1
,
2
where φ ≡
α
1−α
(6)
(7)
> 0 and θt , θt+1 > 0 are utilization parameters associated with the production
in the current period and the storage decision for the next period, respectively.7 In equilibrium,
the aggregate resource and inventory demands are provided as a function of exogenous (rt+1 ,
Nt ),8 and endogenous state variables (It , Qt , zt , ξt , χt ).
Notice that the left-hand side of the inventory demand equation (7) is the convenience
yield, which is the refinery’s expected benefit from holding an additional barrel of crude oil.
The proposed convenience yield is time-varying (Gibson and Schwartz 1990) in line with state
variables; it is inversely related to the current inventories (Pindyck 1994, Alquist and Kilian
2010, and Knittel and Pindyck 2015), indicating that inventory holders have larger benefits
when inventories are relatively low. Furthermore, it exhibits asymmetric responses to the longterm dynamics and short-term deviation components of the crude oil spot process (Schwartz
and Smith 2000). In general, the proposed convenience yield is consistent with the theory of
storage.
However, it is worth noting that the inverse relationship in this paper is derived within
the optimal storage decision of the refinery who naturally has an incentive to hold crude oil
7
For example, θ1 are used for both θt and θt+1 when the period t corresponds to March. However, we use θ2
in place of θt+1 when the period t corresponds to May. It is because it becomes summer in the next period while
it is currently spring.
8
In the aggregate resource demand equation (6), the number of the representative producers in the previous
period (Nt−1 ) works as “positive externality,” yet the number at current period (Nt ) works as “negative externality.” This confirms our intuition under the limited storage facilities; an increasing competition tends to raise
marginal storage costs in the short run due to congestion, but it also motivates growths in storage technologies
in the long run.
12
inventories, which has never been done in the literature. Instead, researchers commonly assume
such a negative relationship when studying the behavior of crude oil price and its relationship
with inventories via convenience yield. When measuring the convenience yield, the proposed
storage model also differs from earlier researchers’ approaches in two ways. First, the proposed
convenience yield reflects convex storage costs that are essential considerations for refiners (or
inventory holders in general) besides their expectations on future price movements. Second, data
on crude oil inventories are used along with estimates for parameters and unobservable state
variables. These contrast with the widely adopted approach of inferrring the net convenience
yield (abstracting marginal storage costs) by using the term spread calculated from prices for
crude oil futures contracts (Pindyck 1994, Alquist and Kilian 2010, and Knittel and Pindyck
2015). In the following section, I provide a historical convenience yield measure calculated from
the inventory demand equation (7), using parameter estimates as well as state variables that
include data on crude oil inventory.
3
3.1
Empirical Results
Data and Empirical Implementation
The monthly global crude oil inventories, consumptions and prices for crude oil futures contracts
are used for estimating the model. Given the lack of data on global crude oil inventories, I construct the monthly series of global crude oil inventories following Kilian and Murphy (2014),
where total U.S. crude oil inventories are scaled by the ratio of Organisation for Economic
Co-operation and Development (OECD) petroleum stocks over U.S. petroleum stocks. Next, I
construct the monthly series of global crude oil consumptions by subtracting monthly changes in
global crude oil inventories from monthly global crude oil productions. While OECD petroleum
stocks is provided by the International Energy Agency, U.S crude oil inventories, petroleum
stocks and global crude oil productions are provided by the U.S. Energy Information Administration (EIA). Then, I contruct monthly prices for the crude oil futures contracts from daily
prices for the light sweet crude oil (aka WTI crude oil) futures contracts traded in the New York
Mercantile Exchange. More specifically, after obtaining daily prices for futures contracts having
maturities up to 13 months from datastream, I construct monthly prices of a given maturity
13
using the first calendar date price of corresponding futures contracts.9 The monthly U.S. LIBOR is used for representing the effective historical risk-free interest rate. Lastly, the U.S.CPI,
provided by OECD, is used for deflating nominal prices of the crude oil futures contracts into
real prices as well as adjusting nominal interest rates for realized inflation. The time period of
the analysis is from March 1989 to November 2014, where the beginning period of the analysis
is determined from the availability of prices for a 12-month crude oil futures contract, and the
ending period is determined from the availability of the OECD Inventory.
Assuming an inelastic supply of crude oil following Kilian (2008) and Alquist and Kilian
(2010),10 I estimate model parameters along the aggregate demands provided by (6) and (7).
As noted earlier, I treat crude oil spot price as being unobservable since the spot price is not
available in general.11 Instead, I use prices of crude oil futures contracts to estimate unobservable components in the crude oil spot price following Schwartz and Smith (2000). Furthermore, I
allow for measurement errors in the observables such as crude oil demands, inventories and prices
for crude oil futures contracts during the estimation procedure. This addresses the widely recognized measurement issues in oil quantity variables (Baumeister and Kilian 2012) and mispriced
oil futures contracts at times (Kilian and Murphy 2014)
Now, I provide three sets of observation equations as linear functions of state variables
(zt , ξt , χt ) for the estimation of the proposed storage model. The first two observation equations
are readily available from equilibrium decisions for the aggregate resource demands and inventory
demands.
Consider the aggregate resource demands from (6). Taking the log and subtracting the
9
In general, the construction of the monthly prices is consistent with the procedure described in Alquist and
Kilian (2010). One exception is to use the price of τ -month futures contract at the first calendar date in this
paper, compared to using the end-of-month price of the same futures contract in Alquist and Kilian (2010). Both
approaches are intended to avoid the potential liquidity issue in the futures market since trading in the current
delivery month ceases on the third business day prior to the 25th calendar day of the month preceeding the
delivery month.
10
Given capacity constraints in world crude oil production, the supply side of the crude oil market has been
commonly modeled by an endowment structure, where unexpected disruption/discovery in the oil production
shifts the supply curve. While estimating parameters along the aggregate demands, I confirm that forecasting
errors in resource and inventory demand equations are not correlated with the exogenous crude oil supply shock
suggested in Kilian (2009).
11
Strictly speaking, the spot price is not available. While using the cash price for proxying the spot price,
Knittel and Pindyck (2015) note the difference between the two; whereas the spot price is a price for immediate
delivery at a specific location of a specific grade of oil, the cash price refers to an average transaction price, which
reflects discounts or premiums resulting from buyer-seller relationships.
14
previous aggregate resource demands (log Qt−1 ) from the resulting equation yields
log (Qt /Qt−1 ) = φzt − (1 + φ) ξt − (1 + φ) χt
(8)
+ (1 + φ) log α (1 − exp (−θt It−1 /Nt−1 )) + log Nt − log Qt−1 .
Similarly, consider the log of the aggregate inventory demands (7) and linear approximations
of non-linear terms in the resulting equation. Linear approximations are taken for log (It )
being around log (It−1 ), and for χt being close to
−λχ
1−kQ
since χt is mean-reverting around
−λχ
.
1−kQ
Rearranging this, we obtain a linear relationship between a change in the aggregate demands
for inventories (It ) and state variables (zt , ξt , χt ) as
log (It /It−1 ) = Φ1 (xt ) · zt + Φ2 (xt ) · ξt + Φ3 (xt ) · χt + Φ4 (xt ) ,
(9)
where coefficients Φ1 (xt ) , Φ2 (xt ) , Φ3 (xt ) , Φ4 (xt ) are functions of parameters and a vector of
observations xt ≡ (rt+1 , It−1 ) that are either exogenous or predetermined at period t. Appendix
A.1. provides detailed derivations as well as coefficients of the above inventory observation
equation (9).
Lastly, I use closed-form spot price forecasts provided in (5) for evaluating prices of crude
oil futures contracts with various maturities. To see this, consider the risk-neutral stochastic
processes in Section 2.4. Under the risk-neutral valuation paradigm, the “no arbitrage” price
of the contingent claims coincides with the expected future cash flows under the risk-neutral
stochastic process. Since there is no cash payment when a futures contract is traded, the price
of the crude oil futures contract with one month to maturity (Ft,1 ) coincides with the one-periodahead spot price forecast under the risk-neutral measure. Similarly, the price of the crude oil
futures contract with τ periods to maturity (Fτ,t ) is the τ -period-ahead spot price forecast under
the risk neutral measure, i.e., Ft,τ = EtQ [St+τ ].12 After taking log on both sides, I subtract the
logarithm of the same crude oil futures contract in the previous period (log Ft−1,τ +1 ). This
12
Hamilton and Wu (2014) shed new light on the time-varying risk premium in oil futures markets, relying
on an affine factor structure to futures prices. In this paper, risk-premium is also time-varying and depends on
a short-term deviation process, which is a factor reflecting the interaction between the level of inventories and
temporal behaviorsof spot prices. To seethis, the risk premium associated with τ -month to maturity crude oil
futures contract is EtQ [St+τ ] − Et [St+τ ] /Et [St+τ ] ≈ log EtQ [St+τ ] − log Et [St+τ ]. Using spot price forecasts
under the risk-neutral
measure (5) and under the
physical measure without risk adjustments, this further reduces
to A k, kQ ; τ · χt + B k, kQ , λξ , λχ , σ2 , σ3 , ρ; τ , a time-varying function of the short-term deviation process (χt ),
as well as parameters through A (·) and B (·), of which signs and magnitudes depend on model parameters.
15
provides a linear relationship between a periodic return of the crude oil futures contracts with
τ periods to maturity and state variables (ξt , χt ) as
log (Ft,τ /Ft−1,τ +1 ) = ξt + k
2
where sQ
τ ≡ τ · σ2 +
1−(kQ )
Q τ
τ
1 − kQ
1
· χt − τ · λξ − λχ ·
+ sQ
− log Ft−1,τ +1 ,
Q
1−k
2 τ
2τ
1−(kQ )2
(10)
τ
· σ32 + 2ρσ2 σ3 ·
1−(kQ )
1−kQ
is the conditional variance associated with
τ -periods-ahead forecast of the crude oil spot price.
Given three sets of observation equations, namely, (8) for crude oil consumption, (9) for
inventories and (10) for futures contract prices for 1, · · · , 12 months to maturity, I estimate
parameters by maximum likelihood using the Kalman filter. Let ζt = (zt , ξt , χt )0 denote a 3 × 1
vector of unobservable state variables and yt denote a 14 × 1 vector of observed variables at
each period t, i.e., yt = (ln Qt /Qt−1 , log (It /It−1 ) , ln Ft,1 /Ft−1,2 , . . . , ln Ft,12 /Ft−1,13 )0 . Then
yt can be described in terms of a linear function of unobservable state vector ζt . The state-space
representation is given by the following system of equations:
ξt+1 = C + Gξt + vt+1 ,
yt = a (xt ) + H (xt )0 ξt + wt ,
where vt+1 =
Q
Q
Q
1,t+1 , 2,t+1 , 3,t+1
0
is a 3 × 1 vector of risk-neutral innovation processes and
wt is a 14 × 1 vector of measurement errors. In the transition equation, C = (0, −λξ , −λχ )0
and G = diag 1, 1, k Q with diag being a diagonal matrix operator whose diagonal elements
are provided in the parentheses. In the observation equation, H (xt )0 and a (xt ) are specified
following (8), (9) and (10).
Given observations (yt , xt ) for t = 1, . . . , T , the sample log-likelihood can be constructed
using the Kalman filter
L (Θ) =
T
X
log fYt |Xt, Y t−1 yt |xt , y t−1 ,
t=1
where fYt |Xt, Y t−1 yt |xt , y t−1 is the conditional likelihood for a 14 × 1 vector of observables
conditional on their lagged values as well as available observations that are either exogenous or
predetermined. Appendix A.2. provides a detailed explanation for calculating the above sample
16
log likelihood using the Kalman filter as well as parametric restrictions for H (xt )0 and a (xt ) in
the observation equation.13
3.2
Results
The second column in Table 1 reports maximum likelihood estimates and asymptotic standard
errors.14 Estimates of the model parameters are consistent with the theoretical restrictions and
historical observations. As for the estimates themselves, several points stand out. First, the
output share (α) estimate indicates that crude oil resources account for approximately 39% of
consumption good production globally. Second, when the real price of crude oil was $99.21 per
barrel and the inventory level was 4,138.71 million barrels per day in January 2008, estimates
for storage cost functions indicate that the marginal storage cost necessary for increasing an
additional barrel of crude oil inventory is approximately $0.91 per barrel, holding other things
being equal. Third, I find smaller estimates of the utilization parameters for spring (θ1 ) and
summer (θ2 ) than other seasons. These estimates indicate that utilizations for spring and summer are higher than those for fall and winter since the smaller θ implies the refinery’s higher
dependence on available inventories. While all seasonal utilization parameters (θ1 , θ2 , θ3 , θ4 )
are statistically significant, the smallest difference among utilization parameters is 0.0009 between fall (θ3 ) and winter (θ4 ).15 When crude oil inventories correspond to the median level of
2,455.15 millions barrels per day during July 2003, a 0.0009 decrease in θ yields a $0.10 per
barrel increase in the convenience yield holding other variables constant. Fourth, positive mean
reversion (k Q ) indicates that the short-term deviation of the crude oil spot price is highly persistent under the risk-neutral measure. Given the slowly moving long-term trend and the positive
correlation (ρQ ) between the two underlying processes, this further implies the highly persistent
crude oil spot process under the risk-neutral measure. Fifth, the long-term risk premium (λξ ) is
positive and statistically significant, indicating a risky long-term trend component in crude oil
price movement. Given the highly persistent oil price, producers need to provide counterparties
13
See more details regarding the state-space representation and the Kalman filter for maximum likelihood
estimation in Hamilton (Time Series Analysis, 1994).
14
To estimate the model, the likelihood function is reparametrized while preserving restrictions on parameters.
The reported MLE are re-calculated from the estimated MLE, and asymptotic standard errors are calculated
based on initial parametric specifications.
15
In general, seasonal utilization parameters are statistically different from each other at 10% significance level
except two consecutive pairs of summer (θ2 ) and fall (θ3 ), and fall (θ3 ) and winter (θ4 ). Wald statistic for testing
the null hypothesis of H0 : θ1 = θ2 = θ3 = θ4 is 5.99 with the corresponding p-value being 0.11.
17
with sufficient compensations for hedging their price risks, especially being associated with the
long-term risk factor. On the other hand, the short-term risk premium (λχ ) is negative, yet
statistically insignificant. Hence, arbitrageurs do not necessarily require a high risk premium
when providing liquidities for hedgers when facing the short-term risk factor.
Figure 1 plots the historical decomposition of the crude oil spot price with the long-term
trend (Panel 1) and the short-term deviation (Panel 2). The model fit is provided in Panel 3
by overlaying the model forecasts from (4) with the real price of the nearest maturity crude oil
futures contract (henceforth, crude oil price) on a logarithm scale. In the historical decomposition, previous episodes in the crude oil market are shown through the lens of the long-term
and short-term processes. As a random walk process, the long-term trend represents a slowly
moving trend in the crude oil spot price. While stable until 2003, the long-term trend has been
increasing from the beginning of 2004 to the first half of 2008, reaching at the higher equilibrium
level compared to the previous period. On the other hand, the short-term deviation process is a
−λ
mean-reverting process around its long-run average ( 1−kχQ = 0.44), of which historical observations are consistent with previous episodes documented in the earlier literature (Hamilton 2009,
Kilian 2009, and Kilian and Murphy 2014).
Panel 1 in Figure 2 plots the equilibrium convenience yield as a percentage of the crude oil
spot price. The convenience yield fluctuates considerably along the business cycle, repeating the
same patterns around the past recession and recovery periods. When the global economy starts
to recover from the early 2000s’ recession, for example, the convenience yield sharply increases
from 21.72% in February 2002 to 24.03% in February 2003. On the other hand, it declines from
18.43% to 16.67% between December 2007 and June 2009 (Great Recession). The procyclical convenience yield can be explained by its negative relationship with crude oil inventories
(Pindyck 1994). During the early 2000 recovery period, the crude oil inventories declined by
0.12%, which reflects increasing crude oil demands for immediate gasoline productions as well
as for future uses. On the other hand, crude oil inventories had increased by 3.87% during the
Great Recession. The correlation coefficient estimates are −0.46 with the level of crude oil inventories and −0.21 with the monthly changes in crude oil inventories. This finding is consistent
with the theory of storage literature. Moreover, I find asymmetric responses of the convenience
yield toward two components in the crude oil spot price. The convenience yield is positively
(negatively) related to the short-term deviation (long-term trend) components with the correla18
tion coefficient being 0.56 (−0.45) over the sample period. When the convenience yield increased
during the early 2000 recovery period, the short-term deviation increased by 95.12%, whereas
the long-term trend decreased by 3.00%. During the Great Recession, the short-term deviation
had decreased by 33.24%, whereas the long-term component had increased by 3.17%. Lastly,
I find that the historical average convenience yield is 19.10% over the sample period. Holding
other factors of production constant, historical average gains in the refinery’s productivity can
be translated into a significant benefit such as 19.10% relative to the spot price at which the
refinery purchases an additional barrel of crude oil for future production. This again highlights
the importance of using data on inventories and taking associated storage costs into account
when measuring convenience yield.
For comparison, Panel 2 plots the negative of crude oil futures spreads, which are widely
adopted convenience yield measures among researchers. Here I extend the spread series in Figure
4 of Alquist and Kilian (2010) using monthly real prices for futures contracts with three, six,
nine and 12-months to maturity and real crude oil spot price provided by the EIA. In general,
observations from negative spreads are analogous to those from the proposed convenience yield
in Panel 1. One exception is the Persian Gulf War period, where the negative spreads exhibit
the positive spike provided by the sharp spikes in prices for futures contracts. Besides the fact
that the size of spikes are larger for the spreads from longer-term futures contracts,16 it can
be explained by the hypothesized role of crude oil inventories, leaving a precautionary role as
unmodelled dynamics. Despite such a difference, one can visually confirm that the proposed
convenience yield exhibits qualitatively similar patterns during historical events, documented in
Alquist and Kilian (2010).
3.3
Stable Forecasting Relationship
Given increased financial investors’ participation since 2004, it has been a controversial issue
that these new participants have changed the crude oil market structure. For example, Tang
and Xiong (2012) find an increased correlation between oil prices and prices of non energy
16
The proposed convenience yield in Panel 1 measures the expected benefit from one-month storage of an
additional barrel of crude oil. From this perspective, the spread to be compared with Panel 1 should be the
shortest spread defined as a (log) real price difference between spot and futures contract with one month to
maturity. When calculating this, I find the small magnitude of the one-month spread to be compared with the
convenience yield. Moreover, for a few periods, it exhibits erroneous behaviors that are qualitatively different
from spreads provided by longer maturity futures contracts.
19
commodities, raising doubts about potential spillover effects on commodity markets from other
financial markets. On the other hand, Fattouh (2010) shows that the increased liquidity provided
by financial investors facilitates arbitrage activities, enhancing oil market integration. In this
section, I begin by providing evidence of a structural change by estimating the model under two
sub-sample periods split after June 2004. Then, I show that recursive forecasts provided by the
proposed model are more accurate than the no-change forecast at three, six, nine and 12-month
horizons, highlighting that the price-inventory relationship is stable despite a structural change
in the crude oil market.
I estimate the proposed model again using data for two sub-sample periods split after June
2004. The first sub-sample spans from March 1989 to June 2004 (184 observations), and the
second sub-sample spans from July 2004 to November 2014 (125 observations). The third and
fourth columns in Table 1 report maximum likelihood estimates and asymptotic standard errors.
In general, estimates are qualitatively similar to each other and those from the total sample.
Turning to the sizes of parameter estimates, however, there is a clear sign of a structural change
in the oil market. Here, I find a reduced role of crude oil inventory relative to timely purchased
resources from an increased convenience yield parameter (θt ) while preserving the same seasonal
pattern. The convenience yields calculated from parameter estimates across two sub-sample
periods exhibit the same piecewise historical movement, yet at different levels. Holding other
factors of production constant across two periods, the average convenience yields are 33.15%
and 15.00%, respectively, for the first and second periods. However, this does not indicate the
lower convenience yield during the latter period due to the omitted factors in the refinery’s
gasoline production.17 Instead, when combined with the increased crude oil inventories, such
a difference is reflected in the model through higher trend (g) and lower marginal storage cost
(c1 ) estimates for the second period. From parameter estimates describing the crude oil spot
price, I find that the statistical properties for the crude oil spot price has also changed. More
specifically, the crude oil spot price becomes more persistent, which is evidenced by the increased
autocorrelation (k Q ) for the mean-reverting short-term process and increased cross-correlation
(ρQ ) between short- and long-term processes. The increased persistence of the crude oil spot
price results in higher risk premiums that are necessary for refiners to hedge heightened risks
17
Similarily, the smaller output share (α) from the second sub-period does not imply the reduced importance
of the crude oil resource in gasoline production.
20
in the crude oil market. This can be confirmed by the increased prices of both long- (λξ ) and
short-term (λχ ) risks during the latter period. To sum up, the sub-sample exercise in this section
provides empirical evidence of the structural change in the crude oil market, which are consistent
with Tang and Xiong (2012), Büyükşahin and Robe (2014) and Hamilton and Wu (2014) among
others.
Next, I perform an out-of-sample forecasting exercise to investigate whether the structural
change affects the price-inventory relation proposed in this paper. Throughout the forecasting
exercise, I forecast the log price of oil because the model specifies the relationship between the log
price and the state variables. To reduce computational burden during the recursive estimation
procedure, I use a smaller set of observations consisting of futures contract prices with one,
three, six and 12 months to maturity, oil consumptions and inventories. Given evidence for the
structural break, I adopt a rolling fixed estimation window method based on its robustness in
the presence of nonstationarity (Giacomini and White 2006). With 184 months of estimation
window, I fit the model to a sample of 184 months, generate one-step-ahead forecasts for prices of
futures contracts with one, three, six and 12 months to maturity and drop the oldest observation
from the sample when adding the new data. I repeat this process and evaluate the performance
of 125 monthly out-of-sample forecasts from July 2004 to November 2014. I also generate one,
three, six and 12-month forecasts for the crude oil spot price by using fitted unobservable state
variables and parameter estimates.18 To evaluate the performance, I set the benchmark as the
random walk model forecast, of which a one-month-ahead forecast for τ -month to maturity
futures contract is the current price for the same futures contract and the τ -month-ahead spot
price forecasts coincides with the current crude oil spot price.
Table 2 reports recursive mean squared prediction error (MSPE) and mean absolute error
(MAE) ratios relative to the no-change forecasts. Panel 1 evaluates performance of one-monthahead price forecasts for crude oil futures contracts with one, three, six and 12 months to
maturity. Compared to the benchmark, the proposed model provides less accurate price forecasts
for all futures contracts using MSPE criteria (column 2). Although more accurate forecasts
18
Let ξt+1|t and χt+1|t be one-month-ahead forecasts for long-term trend and short-term deviation processes,
respectively. Given parameter estimates, price forecasts for τ -month maturity futures contract are calculated
from (10) with unobservable state variables (ξt , χt ) being replaced by their forecasted values (ξt+1|t , χt+1|t ).
Similarily, denoting by ξt|t and χt|t the fitted state variables, the τ -month- ahead forecast for crude oil spot price
is calculated from (5) with unobservable state variables (ξt , χt ) being replaced by their fitted values (ξt|t , χt|t ).
Appendix A.2. provides a detailed explanation for the recursive calculation necessary for generating fitted values
(A4, A5) and forecasts (A6, A7) of unobservable state variables.
21
for short-maturity futures contracts are found from the MAE criteria (column 6), I also find
that such improvements are not statistically significant based on tests for equal forecasting
accuracy suggested by Diebold and Mariano (1995) and West (1996) (column 7). Given abnormal
behaviors in the financial markets during the Great Recession period (Gürkaynak and Wright
2012), I recalculated the relative MSPE and MAE ratios after ruling out price forecasts during
this period. When excluding forecasts from November 2008 to March 2009, I find that the
model provides more accurate one-month-ahead price forecasts for contracts with one and three
months to maturity under MSPE (column 4) and one, three, and six months to maturity under
MAE (column 8). Yet, the statistical significance for such improvements is weak in general.
These results indicate that prices for futures contracts are harder to predict than spot prices.
A better forecast of the futures price could come from explicitly modeling the risk premium in
these contracts as in Baumeister and Kilian (2014).
Panel 2 reports recursive MSPE and MAE of one, three, six and 12-month-ahead crude oil
spot price forecasts relative to the no-change forecast for crude oil spot prices. Compared to
the no-change forecast, I find that the model provides more accurate spot price forecasts for all
horizons considered under both MSPE and MAE criteria. Such improvements are statistically
significant at the 10% level for three and 12-month-ahead forecasts using MSPE (columns 2
and 3), and for three, six and 12-month-ahead forecasts using MAE (columns 6 and 7). When
ruling out a few forecasts for the evaluation, I find larger and more statistically significant
improvements across forecasting horizons in general under both MSPE and MAE criteria. These
findings indicate that a price-inventory relationship proposed in this paper is stable despite
potential structural changes in the oil market, consistent with Kilian and Murphy (2014) who
develop a four-variable structural VAR model. The improved forecasting accuracies in predicting
future crude oil spot price are consistent with Baumeister and Kilian (2012) who document the
improved accuracies from widely adopted forecasting models when using real-time data.
4
The Effect of Financial Investors in the Crude Oil Market
So far, I investigate the role of crude oil inventories for the refinery’s gasoline production and
document empirical evidence for the convenience yield around which crude oil inventories and
prices center. Although failing to predict prices of crude oil futures contract, I show that the
22
proposed storage model provides more accurate crude oil spot price forecasts than the no-change
model.
In this section, I illustrate one use of the framework by re-examining the possible contribution of financial investors using the proposed model. To do so, I begin by introducing simple
forecasting models that also use oil inventories as an explanatory variable. Then I investigate a
breakdown in their forecasting relationships since 2004, which contrasts with the stable priceinventory relationship documented in the previous section.
Traditionally, traders would react to releases of inventory data in a predictable way. Unexpectedly high inventories were taken as a sign of weak demand and strong supplies, and the
price would decline. Unexpectedly low inventories would predict an increase in the price. The
fact that this relationship seemed less reliable in the period of heavy financial participation led
some to conclude that financial investors were changing the fundamental behavior of crude oil
prices.
One description of the traditional relation between prices and inventories was developed by
Ye et al. (2002), who defined the normal inventory, N It as the fitted value from a regression of the
observed OECD inventory OIt on a constant, deterministic time trend and monthly dummies:
OIt = β0 + β1 · t +
12
X
βp · Dp ,
(11)
p=2
where (Dp , p = 2, . . . , 12) are monthly dummies; i.e., Dp = 1 if the period t corresponds to the
month p, but Dp = 0 otherwise. They further defined RIt ≡ OIt − N It as the residual from this
regression. Allowing for the possibility that positive and negative residuals had different effects
as well as an effect of the year-over-year change AIt = OIt − OIt−12 , Ye et al. (2002) proposed
a forecasting model (hereafter, traditional forecasting model) as follows:
F1,t = a +
5
X
i=0
bi · RIt−i +
5
X
ci · LIt−i + d · AIt + e · F1,t−1 + εt ,
(12)
i=0
where F1,t is a real price of the nearest-maturity crude oil futures contract in month t and εt is
a regression error.
Figure 3 plots in-sample and out-of-sample forecasts following the traditional forecasting
model. The period for in-sample forecasts coincides with the earlier literature from January
23
1992 to February 2001. Using estimated coefficients from the in-sample forecasting period, I
calculate out-of-sample forecasts from March 2001 to November 2014. The solid line in March
2001 denotes the beginning of out-of-sample forecasts. The long-dashed line in Figure 3 confirms
the strong in-sample forecasting ability documented in the earlier literature. However, its strong
forecasting ability disappears over the out-of-sample forecasting period.
Given a poor performance of the traditional forecasting model after 2004, Merino and Ortiz (2005) proposed the extended forecasting model by building upon the former model. After
expressing RIt in a percentage of total level (RI%t ) and including U.S. Commodity Futures Trading Commission (CFTC) non-commercial long position (Xt ), the extended forecasting model is
provided as follows:
F1,t = β0 + β1 · t + β2 · ∆RI%t + β3 · RI%t−1 + β4 · ∆Xt + β5 · Xt−1 + ηt ,
(13)
where ∆RI%t = RI%t − RI%t−1 , ∆Xt = Xt − Xt−1 and ηt is a regression error. One thing to
note is that Merino and Ortiz (2005) included the deterministic time trend (t) in the forecasting
equation as opposed to the lagged price (F1,t−1 ) considered in Ye et al. (2002). Figure 4 plots
in-sample and out-of-sample forecasts following the extended forecasting model. The period for
in-sample forecasts coincides with Merino and Ortiz (2005) from January 1992 to June 2004.
Using estimated coefficients from the in-sample forecasting period, I calculate out-of-sample
forecasts from July 2004 to November 2014. The solid line in July 2004 denotes the beginning
of the out-of-sample forecasts. Along with the extended model forecasts, I also calculate the
basic model forecasts from the literature, where the latter model excludes Xt in the former
forecasting equation (13). Figure 4 confirms the improved in-sample forecasting ability of the
extended model as documented in Merino and Ortiz (2005): The extended model exhibits the
better in-sample forecasts than the basic model from 2001 to 2004, while both generate similar
in-sample forecasts from 1992 to 2001. However, when I extend the sample to include data since
publication of Merino and Ortiz’s study, even their model with non-commercial traders does
quite badly. This finding contrasts with the stable price-inventory relationship evidenced by the
improved forecasting accuracies in the previous section.
On the other hand, the optimizing model developed in Section (2) suggests that both equations (12) and (13) are misspecified. To see whether the proposed model exhibits any of the
24
instabilties of either (12) and (13), I perform an analogous out-of-sample evaluation. In order to
highlight the stability of the relationship, I use the same parameter estimates from pre-June 2004
in Table 1 for calculating out-of-sample forecasts recursively from July 2004 to November 2014.
Given fixed parameter estimates, I iterate the following three steps recursively for calculating
one-month-ahead oil price forecasts while updating available observations. First, given available
observations, I update the state vector and the variance matrix using the formula for updating a linear projection. Second, given updated vectors and matrices, I calculate one-monthahead forecasts with linear projections. Third, given one-month-ahead forecasts, I calculate
one-month-ahead out-of-sample forecasts. Figure 5 displays forecasts of the nearest-maturity
crude oil futures contract price. The solid vertical line in July 2004 denotes the beginning of the
out-of-sample forecasts. I find the stable crude oil price forecasting relationship both in-sample
and out-of-sample. In this forecasting exercise, out-of-sample forecasts are based on information
that is either known in advance or predetermined at each forecasting period. Hence, I find that
currently available information is sufficient to produce a reliable forecast for the next month’s
crude oil futures price given the stable price-inventory relationship proposed in this paper.
If equation (12) was misspecified, why did it seem to have such a good fit prior to 2001? To
answer this question, suppose that inventories had been exactly as predicted by the proposed
c t , the predicted equilibrium level
model; that is, replace OIt on the left-hand side of (11) with OI
of inventories from equation (9). I then reproduce the Ye et al. exercise on this generated data,
the results of which are shown in Figure 3 (short-dashed line). Even though equation (9) is the
correct relationship that exactly describes these generated data, and even though (11) and (12)
are misspecified, they appear to do a reasonable job until 2004, after which they completely
break down.
The explanation for the apparent success of the misspecified model in the earlier period can
be found in the fact that the permanent component of crude oil prices, ξt , was relatively stable
until 2003 but then began to increase substantially (see the top panel of Figure 1). During
the period when ξt was relatively stable, simple relationships like (11) did not do a bad job of
fitting the data, even though they fundamentally mischaracterize the true relationship between
inventories and crude oil prices. With the large changes in ξt observed since 2004, the need for
a better model of the relationship between oil prices and inventories, such as that proposed in
this paper, becomes quite evident.
25
5
Conclusion
This paper proposes an equilibrium storage model of the global crude oil market to study the
role of crude oil inventories in refiners’ gasoline production. Given a heavy dependence of
refinery’s production on resources on hand, I model the role of crude oil inventories directly
as enhancing gasoline production by treating inventories as an essential factor of production,
following Kydland and Prescott (1982, 1988) and Ramey (1989). The model is simple, yet it
provides a measure for convenience yield that refiners expect to receive when storing a barrel of
crude oil.
Using data on crude oil inventories, consumption and prices for futures contracts, I find
empirical evidence on convenience yield that is consistent with the theory of storage literature.
I show that the proposed convenience yield is inversely related to inventories, yet positively
related to oil prices. While exhibiting seasonal and procyclical behaviors, I show that the historical convenience yield is about 19% of the oil price on average from March 1989 to November
2014. Moreover, I identify a structural change in crude oil market fundamentals since 2004
that coincides with the period of rapidly increasing financial investors’ activities. I document
empirical evidence on the structural change from a rising long-term component in the spot price
process and from changing parameter estimates for before and after June 2004. Despite the
structural break, I show that the price-inventory relationship proposed in this paper is stable
over time, providing more accurate forecasts for future crude oil prices compared with the random walk model as well as earlier forecasting models. In light of this evidence, I conclude that
the contribution of financial investors’ activities is weak in the crude oil market.
26
References
Acharya, V. V., Lochstoer, L. A., and Ramadorai, T. (2013). Limits to arbitrage and hedging: evidence
from commodity markets. Journal of Financial Economics, 109(2):441–465.
Alquist, R. and Kilian, L. (2010). What do we learn from the price of crude oil futures? Journal of
Applied Econometrics, 25(4):539–573.
Basak, S. and Pavlova, A. (2015). A model of financialization of commodities. Journal of Finance.
Forthcoming.
Baumeister, C. and Kilian, L. (2012). Real-time forecasts of the real price of oil. Journal of Business &
Economic Statistics, 30(2):326–336.
Baumeister, C. and Kilian, L. (2014). A general approach to recovering market expectations from futures
prices with an application to crude oil.
Brennan, M. J. (1958). The supply of storage. American Economic Review, 48(1):50–72.
Bryant, H. L., Bessler, D. A., and Haigh, M. S. (2006). Causality in futures markets. Journal of Futures
Markets, 26(11):1039–1057.
Büyükşahin, B. and Robe, M. A. (2014). Speculators, commodities and cross-market linkages. Journal
of International Money and Finance, 42:38–70.
Considine, T. J. (1997). Inventories under joint production: an empirical analysis of petroleum refining.
Review of Economics and Statistics, 79(3):493–502.
Deaton, A. and Laroque, G. (1992). On the behaviour of commodity prices. Review of Economic Studies,
59(1):1–23.
Diebold, F. X. and Mariano, R. S. (1995). Comparing predictive accuracy. Journal of Business and
Economic Statistics, 13(3):253–265.
Duffie, D. (2001). Dynamic asset pricing theory. Princeton University Press.
Eckstein, Z. and Eichenbaum, M. S. (1985). Inventories and quantity-constrained equilibria in regulated
markets: the us petroleum industry, 1947-1972. Energy, Foresight, and Strategy.
Fama, E. F. and French, K. R. (1987). Commodity futures prices: some evidence on forecast power,
premiums, and the theory of storage. Journal of Business, 60(1):55–73.
Fama, E. F. and French, K. R. (1988). Business cycles and the behavior of metals prices. Journal of
Finance, 43(5):1075–1093.
Fattouh, B. (2010). The dynamics of crude oil price differentials. Energy Economics, 32(2):334–342.
Fattouh, B., Kilian, L., and Mahadeva, L. (2013). The role of speculation in oil markets: what have we
learned so far? Energy Journal, 34(3):7–33.
Giacomini, R. and White, H. (2006). Tests of conditional predictive ability. Econometrica, 74(6):1545–
1578.
Gibson, R. and Schwartz, E. S. (1990). Stochastic convenience yield and the pricing of oil contingent
claims. Journal of Finance, 45(3):959–976.
Gorton, G. B., Hayashi, F., and Rouwenhorst, K. G. (2013). The fundamentals of commodity futures
returns. Review of Finance, 17(1):35–105.
Gürkaynak, R. S. and Wright, J. H. (2012). Macroeconomics and the term structure. Journal of Economic
Literature, 50(2):331–367.
27
Hamilton, J. D. (1994). Time series analysis. Princeton University Press.
Hamilton, J. D. (2009). Causes and consequences of the oil shock of 2007–08. Brookings Papers on
Economic Activity, 1(Spring):215–284.
Hamilton, J. D. and Wu, J. C. (2014). Risk premia in crude oil futures prices. Journal of International
Money and Finance, 42:9–37.
Hamilton, J. D. and Wu, J. C. (2015). Effects of index-fund investing on commodity futures prices.
International Economic Review, 56(1):187–205.
Irwin, S. H. and Sanders, D. R. (2011). Index funds, financialization, and commodity futures markets.
Applied Economic Perspectives and Policy, 33:1–31.
Juvenal, L. and Petrella, I. (2015). Speculation in the oil market. Journal of Applied Econometrics,
30:621–649.
Kaldor, N. (1939). Speculation and economic stability. Review of Economic Studies, 7(1):1–27.
Kennedy, J. P. (2012). The high cost of gambling on oil. New York Times, April 10.
Kilian, L. (2008). Exogenous oil supply shocks: how big are they and how much do they matter for the
us economy? Review of Economics and Statistics, 90(2):216–240.
Kilian, L. (2009). Not all oil price shocks are alike: disentangling demand and supply shocks in the crude
oil market. American Economic Review, 99(3):1053–1069.
Kilian, L. and Lee, T. K. (2014). Quantifying the speculative component in the real price of oil: The role
of global oil inventories. Journal of International Money and Finance, 42:71–87.
Kilian, L. and Murphy, D. P. (2014). The role of inventories and speculative trading in the global market
for crude oil. Journal of Applied Econometrics, 29(3):454–478.
Knittel, C. R. and Pindyck, R. S. (2015). The simple economics of commodity price speculation. American
Economic Journal. forthcoming.
Kydland, F. E. and Prescott, E. C. (1982). Time to build and aggregate fluctuations. Econometrica,
50(6):1345–1370.
Kydland, F. E. and Prescott, E. C. (1988). The workweek of capital and its cyclical implications. Journal
of Monetary Economics, 21(2-3):343–360.
Masters, M. W. (2008). Testimony before committee on homeland security and governmental affairs. US
Senate, Washington, May 20.
Merino, A. and Ortiz, A. (2005). Explaining the so-called ‘price premium’ in oil markets. OPEC review,
29(2):133–152.
Ng, V. K. and Pirrong, S. C. (1994). Fundamentals and volatility: storage, spreads, and the dynamics of
metals prices. Journal of Business, 67(2):203–230.
Pindyck, R. S. (1994). Inventories and the short-run dynamics of commodity prices. Rand Journal of
Economics, 25(1):141–159.
Pindyck, R. S. (2001). The dynamics of commodity spot and futures markets: a primer. Energy Journal,
22(3):1–30.
Pirrong, C. (2011). Commodity price dynamics: a structural approach. Cambridge University Press.
Ramey, V. A. (1989). Inventories as factors of production and economic fluctuations. American Economic
Review, 79(3):338–354.
28
Routledge, B. R., Seppi, D. J., and Spatt, C. S. (2000). Equilibrium forward curves for commodities.
Journal of Finance, 55(3):1297–1338.
Schwartz, E. and Smith, J. E. (2000). Short-term variations and long-term dynamics in commodity prices.
Management Science, 46(7):893–911.
Smith, J. L. (2009). World oil: market or mayhem? Journal of Economic Perspectives, 23(3):145–164.
Tang, K. and Xiong, W. (2012). Index investment and financialization of commodities. Financial Analyst
Journal, 68(6):54–74.
West, K. D. (1996). Asymptotic inference about predictive ability. Econometrica, 64(5):1067–1084.
Working, H. (1949). The theory of price of storage. American Economic Review, 39(6):1254–1262.
Ye, M., Zyren, J., and Shore, J. (2002). Forecasting crude oil spot price using oecd petroleum inventory
levels. International Advances in Economic Research, 8(4):324–333.
29
Figure 1: Historical Decomposition of the Crude Oil Spot Price
Figure 1 plots the historical decomposition of the crude oil spot price with the long-term trend process
(Panel 1) and the short-term deviation process (Panel 2). Panel 3 plots the crude oil spot price (shortdashed line) together with the price for nearest-maturity crude oil futures contract (solid line) on a log
scale. The unit is dollars per barrel (2010.5=100) on a log scale.
30
Figure 2: Convenience yield
Figure 2 plots the convenience yield measure. Panel 1 displays the equilibrium convenience yield as a
percentage of the crude oil spot price. Panel 2 displays negative spreads calculated from prices of crude
oil futures contracts with 3-, 6-, 9- and 12-months to maturity, and the real crude oil spot price provided
by the U.S. Energy Information Administration.
Figure 3: Forecasts of the Ye et al. (2002) Model
Figure 3 plots in-sample and out-of-sample forecasts of the real crude oil futures price following Ye
et al. (2002). In-sample forecasts are obtained by fitting the traditional forecasting model of Ye et al.
(2002) from January 1992 to February 2001. Out-of-sample forecasts are calculated recursively from
March 2001 to November 2014, using coefficient estimates from the in-sample forecasting period. The
traditional model’s forecasts, Forecasts (T ), are marked by the long-dashed line. The short-dashed line
indicates in-sample and out-of-sample forecasts from the same model, with observed crude oil inventories
being replaced by those predicted from the proposed storage model, which is denoted by Forecasts (R).
For comparison, the solid line plots observed real crude oil futures prices. The vertical line denotes the
beginning of out-of-sample forecasts in March 2001.
31
Figure 4: Forecasts of the Merino and Ortiz (2005) Model
Figure 4 plots in-sample and out-of-sample forecasts of the real crude oil futures price following Merino
and Ortiz (2005). In-sample forecasts are obtained by fitting the forecasting models of Merino and Ortiz
(2005) from January 1992 to June 2004. Out-of-sample forecasts are calculated recursively from July
2004 to November 2014, using coefficient estimates from in-sample forecasting period. The extended
(short-dashed line) and the basic (long-dashed line) refer to the forecasting models with and without
a role for non-commercial traders. For comparison, the solid line plots observed real crude oil futures
prices. The vertical line in July 2004 denotes the beginning of the out-of-sample forecasts.
Figure 5: Forecasts of the Proposed Storage Model
Figure 5 plots in-sample and out-of-sample forecasts of the real crude oil futures price from the proposed
storage model, Forecasts (S ). In-sample forecasts are calculated from the proposed storage model by
using observations from March 1989 to June 2004, and out-of-sample forecasts are calculated recursively
from forecasts for state variables by using parameter estimates from the in-sample estimation period.
32
Table 1: Parameter Estimates
Period
Output share
α
Convenience yield
θ1 (spring)
θ2 (summer)
θ3 (fall)
θ4 (winter)
Storage cost
c1
Trend
g
Risk premium
λξ (long)
λχ (short)
Spot process
kQ
ρQ
σ1
σ2
σ3
Likelihood
1989.3-2014.11
1989.3-2004.6
2004.7-2014.11
0.3933
(0.0122)
0.4347
(0.0154)
0.3470
(0.0118)
0.1854
0.1878
0.1896
0.1905
(0.0059)
(0.0058)
(0.0058)
(0.0059)
0.1397
0.1415
0.1434
0.1436
(0.0078)
(0.0077)
(0.0077)
(0.0077)
0.1761
0.1788
0.1811
0.1822
(0.0074)
(0.0073)
(0.0072)
(0.0074)
0.0051
(0.0003)
0.0152
(0.0017)
0.0063
(0.0008)
0.0003
(< 0.0000)
0.0002
(< 0.0000)
0.0004
(< 0.0000)
0.0026
−0.0443
(0.0002)
(0.0697)
0.0001
−0.0386
(0.0004)
(0.0655)
0.0063
−0.0351
(0.0004)
(0.0486)
0.9014
(0.0018)
0.0968
(0.0596)
0.2470
(0.0126)
0.0518
(0.0021)
0.0789
(0.0033)
14, 191.74
0.9074
(0.0028)
−0.0173
(0.0843)
0.2246
(0.0143)
0.0377
(0.0020)
0.0910
(0.0050)
8, 174.74
0.9313
(0.0021)
0.1206
(0.0966)
0.2638
(0.0190)
0.0652
(0.0042)
0.0623
(0.0045)
6, 852.38
This table reports maximum likelihood estimates (MLE) for the model parameters and their asymptotic
standard errors in parentheses. In the estimation procedure, the likelihood function is reparametrized
while preserving restrictions on parameters. The reported MLE are recalculated from the estimated
MLE, and asymptotic standard errors are estimated by approximating the second derivative of the loglikelihood functions at MLE estimates. The second column provides parameter estimates for the entire
sample period, while the third and last columns report estimates from two sub-samples. In order to
preserve space, estimates for forecasting errors are not reported, although they are jointly estimated with
other model parameters. The last row reports the maximum value achieved for the log of the likelihood
function.
33
Table 2: Recursive MSPE and MAE Ratios Relative to the No-Change Forecast
MSPE
MAE
Total
Excluded
Total
Excluded
Panel 1: Prices of futures contracts with τ -months to maturity
Horizon
1
1.0028 (n/a) 0.9830 (0.33) 0.9875 (0.25) 0.9755 (0.12)
3
1.0044 (n/a) 0.9881 (0.31) 0.9954 (0.36) 0.9847 (0.14)
6
1.0241 (n/a) 1.0076 (n/a) 1.0063 (n/a) 0.9976 (0.43)
12
1.1120 (n/a) 1.1329 (n/a) 1.0421 (n/a) 1.0419 (n/a)
Panel 2: τ -month-ahead spot price forecast
Horizon
1
0.9715 (0.17) 0.9734 (0.23)
3
0.9479 (0.09) 0.8826 (0.07)
6
0.9253 (0.11) 0.8394 (0.10)
12
0.8140 (0.03) 0.7936 (0.05)
0.9806
0.9555
0.9310
0.8776
(0.18)
(0.03)
(0.05)
(0.02)
0.9767
0.9395
0.9126
0.8705
(0.16)
(0.02)
(0.05)
(0.04)
Panel 1 reports the Mean Squared Prediction Error (MSPE) and Mean Absolute Error (MAE) of onemonth-ahead forecasts for prices of futures contracts with 3-, 6-, 9- and 12-months maturities. Panel
2 reports those of 3-, 6-, 9- and 12-month-ahead spot price forecasts recursively provided by the model
forecasts. For each horizon, columns 2 and 6 report recursive MSPE and MAE ratios relative to those
from no-change forecasts. While boldface indicates improvements relative to the no-change forecast,
adjacent columns in parentheses report p-values from tests for equal forecasting accuracy suggested by
Diebold and Mariano (1995) and West (1996). Given abnormal financial market behaviors after the Great
Recession, columns 4 and 8 provide recursive MSPE and MAE ratios calculated after excluding forecasts
for 5 months from 2008.11 to 2009.3.
34
Appendix A: Detailed Explanations
A.1. Equilibrium Prediction and Aggregate Inventory Demands
The equilibrium prediction for the aggregate inventory demands is obtained by combining the representative producer’s optimal decisions with his forecasts for the future crude oil spot price. Recall the
optimality conditions from the model
zt
e Qt =
e
−rt+1
EtQ
1
1−α
It−1
− 1
αe
(St ) 1−α · Nt ,
1 − exp −θt
Nt−1
zt
It
It
α
1−α θt+1
zt+1
Qt+1 ) (Nt+1 )
(e
exp −θt+1
= St 1 + c1
− e−rt+1 EtQ [St+1 ] ,
Nt
Nt
Nt
where we rearrange the equation for aggregate resource demands in (6).
After moving the time subscript one period forward for the rearranged aggregate resource demands,
we plug the resulting equation into aggregate inventory demands whose left-hand side becomes:
h
i
α
It
− α
L (It , Nt , rt+1 ) · EtQ (ezt+1 ) 1−α (St+1 ) 1−α = St 1 + c1
− e−rt+1 EtQ [St+1 ] ,
Nt
α
1−α
α
exp −θt+1 NItt does not dewhere L (It , Nt , rt+1 ) ≡ e−rt+1 α 1−α θt+1 (1 + g) 1 − exp −θt+1 NItt
pend on the representative producer’s forecasts on the future technology process zt+1 and the future spot
price St+1 . Assuming the independence of the technology process and the log-normal property such as
h
i
φ
α
EtQ (ezt+1 ) = exp φzt + 12 φ2 σ12 for φ ≡ 1−α
under the risk-neutral measure, the equilibrium storage
equation is further reduced to:
1 L (It , Nt , rt+1 ) · exp −φ −zt + ξt + k χt − λξ − λχ + φ2 σ12 + sQ
− rt+1
1
2
c1 It
1
= St 1 +
− exp − 1 − k Q χt − λξ − λχ + sQ
− rt+1 .
Nt
2 1
Q
After taking the log of both sides, the measurement equation of the aggregate inventory demands is
obtained by approximating non-linear terms in the resulting equation. Linear approximations are taken
where χt and It are close to
−λχ
1−kQ
and It−1 , respectively, since χt is mean-reverting around
−λχ
1−kQ
and It
is close to It−1 for most of the sample. Denoting by xt ≡ (rt+1 , It−1 ) a vector of obervations that are
either exogenous or predetermined at period t, we have:
log (It /It−1 ) = Φ1 (xt ) · zt + Φ2 (xt ) · ξt + Φ3 (xt ) · χt + Φ4 (xt ) ,
35
where Φ1 (xt ), Φ2 (xt ), Φ3 (xt ) and Φ4 (xt ) are as specified as follows:
−1

θt+1 It−1


φθ
exp
−
t+1
N
c
I
t
t−1
1
−θt+1 −

Φ0 (xt ) = 
It−1

Nt  1− exp − θt+1 It−1
− exp −λξ + 12 sQ
1+ c1N
1 −r t+1
Nt
t

Φ1 (xt ) = Φ0 (xt ) × (−φ)
Φ2 (xt ) = Φ0 (xt ) × (1 + φ)




1 − k Q exp −λξ + 21 sQ
−r
t+1
1
Φ3 (xt ) = Φ0 (xt ) × 1 + φk Q +
It−1


− exp −λξ + 12 sQ
1+ c1N
1 −r t+1
t
n
o
θt+1 It−1
θt+1 It−1
Φ4 (xt ) = Φ0 (xt ) ×
−φ ln 1− exp −
− ln αφ θt+1 (1 + g)
Nt
Nt
1 Q
λ
exp
−λ
+
s
−r
χ
t+1
ξ 2 1
1
c1 It−1
− exp −λξ + sQ
+
+ ln 1+
1 −r t+1
It−1
1 Q
Nt
2
−
exp
−λ
+
1+ c1N
s
−r
t+1
ξ
2 1
t
1
2
−φ (λξ +λχ ) − φ2 sQ
1 +σ 1 +r t+1 .
2
A.2. State-Space Representation and Kalman Filter
Given the set of observations that are linear functions of state variables, I estimate associated parameters
by maximum likelihood using the Kalman filter. First, I represent the system of equations in the state0
space form. Let ζt = (zt , ξt , χt ) denote a 3 × 1 vector of unobservable state variables and yt denote
a 14 × 1 vector of observed variables at each period t. Then yt can be described in terms of a linear
function of unobservable state vector ζt . Denoting by xt a vector of observations that are either exogenous
or predetermined, the state-space representation is given by the following system of equations:
ξt+1 = C + Gξt + vt+1 ,
(A1)
0
yt = a (xt ) + H (xt ) ξt + wt ,
(A2)
where a (xt ) and H (xt ) in the observation equation denote a 14 × 1 vector valued function and a 14 × 2
matrix valued function, which are specified below. In the transition equation, C and G denote respectively a 3 × 1 coefficient vector and a 3 × 3 coefficient matrix; that is, C = 0, −λξ , −λχ
0
and
G = diag 1, 1, k Q with diag being a diagonal matrix operator whose diagonal elements are provided
in the parentheses
0
0
0
Let y t−1 denote data observed through date t − 1; i.e., y t−1 ≡ yt−1
, yt−2
, . . . , y10 , x0t−1 , . . . , x01 .
36
0
Assuming that conditional on xt and y t−1 , the vector vt+1
, wt0

vt+1



| xt ,y t−1  ∼ N 
wt
0
0
 
has the normal distribution
Q 0
,
0
0

 ,
(A3)
R
where Q (i, i) = σ 2i for i = 1, 2, 3 and Q (2, 3) = Q (3, 2) = ρQ σ2 σ3 with σ1 , σ2 , σ3 > 0 and ρQ ∈
2 , σ2 , σ2 , . . . , σ2
(−1, 1) are unknown parameters in state variables. R = diag σQ
I
F,1
F,12
is a 14 × 14
2
measurement error matrix. Here σI2 , σQ
> 0 are measurement errors for aggregate demands for inventories
2
and resources, and σF,τ
is the measurement error for the price of the crude oil futures contract with the
τ -periods until maturity for τ = 1, . . . , 12.
0
We are left with specifying one unknown matrix H (xt ) and one unknown vector a (xt ), yet restrictions
are obtained from three sets of observation equations derived earlier. Recall these observation equations
following (8), (9) and (10). With these, we can specify H (xt ) and a (xt ) following






0
H (xt ) = 





φ
−1 − φ −1 − φ
Φ1 (xt )
Φ2 (xt )
0
..
.
1
..
.
0
1








Φ3 (xt )





a
(x
)
=
Ω1,1 (xt ) 
t




..




.


Ω1,12 (xt )
Ψ (xt )




Φ4 (xt )


Ω2,1 (xt ) 
 ,

..


.

Ω2,12 (xt )
which completes the state-space representation.
Next, consider the calculation of the conditional likelihood using the Kalman filter for the maximum
likelihood estimation. Suppose ξt |y t−1 ∼ N ξbt|t−1 , Pt|t−1 . Since xt contains only strictly exogenous
variables or lagged values of y, we also have ξt |y t−1 , xt ∼ N ξbt|t−1 , P t|t−1 .
With (A1), (A2) and (A3), it can be shown that

ξt



| y t−1 , xt  ∼ N 
yt
ξbt|t−1
0
a (xt ) + [H (xt )] ξbt|t−1
 
,
Pt|t−1
Pt|t−1 H (xt )
0
H (xt ) Pt|t−1
0

 ,
H (xt ) Pt|t−1 H (xt ) + R
where a (xt ) and H (xt ) are treated as deterministic conditional on xt .
Since they are jointly normally distributed, the conditional distribution of ξt given y t−1 , xt , yt becomes
b
ξt |y t−1 , xt , y t ∼ N ξt|t , P t|t where we used the formula for updating a linear projection:
i
−1 h
0
0
ξbt|t = ξbt|t−1 + Pt|t−1 H (xt ) H (xt ) Pt|t−1 H (xt ) + R
× yt − a (xt ) − [H (xt )] ξbt|t−1
−1
0
0
Pt|t = Pt|t−1 − Pt|t−1 H (xt ) H (xt ) Pt|t−1 H (xt ) + R
H (xt ) Pt|t−1 .
37
(A4)
(A5)
One can calculate sample likelihood using the Kalman filter as derived earlier given ξt+1 |y t ∼ N ξbt+1|t , P t+1|t
with
ξbt+1|t = C + Gξbt|t
(A6)
Pt+1|t = GPt|t G0 + Q
(A7)
Assume that the initial state ξ1 is distributed as N ξb1|0 , P1|0 . Given observations {yt , xt } for
t = 1, . . . , T , the distribution of yt conditional on (yt−1 , . . . , y1 , xt , xt−1 , . . . , x1 ) is normal with mean
0
0
a (xt ) + H (xt ) ξbt|t−1 and variance H (xt ) Pt|t−1 H (xt ) + R ; that is, for t = 1, 2, . . . , T , we have the
conditional likelihood following
fYt |Xt, Yt−1 yt |xt , y t−1
=
H (xt )0 Pt|t−1 H (xt ) + R−1/2
0
1
0
yt − a (xt ) − H (xt ) ξbt|t−1
× exp −
2
i
−1 0
0
× H (xt ) Pt|t−1 H (xt ) + R
yt − a (xt ) − H (xt ) ξbt|t−1 .
−n/2
(2π)
With this, the sample log likelihood becomes
L (Θ) =
T
X
log fYt |Xt, Y t−1 yt |xt , y t−1 ,
t=1
where the unknown parameters are estimated numerically.
38